 
Summary: Bayesian Regularization of the Video Matting Problem
Nicholas Apostoloff and Andrew W. Fitzgibbon
Department of Engineering Science, University of Oxford
{nema,awf}@robots.ox.ac.uk
Objective
To regularize the inverse problem of Video Matting in a Bayesian
framework using priors on the distribution of alpha values and the
spatiotemporal consistency of image sequences.
Video Matting
Video matting is a classic image processing problem involving the extraction of a
foreground object from an arbitrary background in a sequence of images. It is most
prevalent in the film industry for special effects shots that require the superposition
of an actor onto a new background.
The compositing equation linearly combines a background image B with a fore
ground image F to form the composite image C using the alpha matte :
C = F + (1  )B
= × + ×
C F (1  ) B
However, the video matting problem is the inverse of this:
Given an sequence of images C, solve for , F and B.
